Re: [RFC PATCH v2, part3 07/11] PCI, xen-pcifront: use new PCI interfaces to simplify implementation

[Jiang Liu <liuj97@gmail.com>]

On Sat 08 Jun 2013 01:07:06 AM CST, Konrad Rzeszutek Wilk wrote:
> On Sat, Jun 08, 2013 at 12:50:31AM +0800, Jiang Liu wrote:
>> On 06/07/2013 11:38 PM, Konrad Rzeszutek Wilk wrote:
>>> On Fri, Jun 07, 2013 at 10:50:24AM -0400, Konrad Rzeszutek Wilk wrote:
>>>> On Thu, May 16, 2013 at 11:50:55PM +0800, Jiang Liu wrote:
>>>>> Use new PCI interfaces to simplify xen-pcifront implementation:
>>>>> 1) Use pci_create_root_bus() instead of pci_scan_bus_parented()
>>>>>    because pci_scan_bus_parented() is marked as __deprecated.This
>>>>>    also gets rid of a duplicated call of pci_bus_start_devices().
>>>>> 2) Use pci_stop_root_bus() and pci_remove_root_bus() instead of
>>>>>    open-coded private implementation.
>>>>> 3) Use pci_set_host_bridge_release() to release data structures
>>>>>    associated with PCI root buses.
>>>>> 4) Use pci_bus_get()/pci_bus_put() to manage PCI root bus reference
>>>>>    count.
>>>>>
>>>>> This is also a preparation for coming PCI bus lock enhancement.
>>>
>>> With this patch from :
>>>
>>>  Merge branch 'pci_lock_v3' of https://github.com/jiangliu/linux into testing
>>>
>>>
>>> it blows up when detaching the device.
>> Hi Konrad,
>>    Thanks for testing! According to the log messages, this issue should
>> be related to pci bus reference counter management. Seems we have done
>> an extra(unbalanced) release of pci bus device.
>>    Will investigate it tomorrow!
>
> That is quite commendable that you are willing to look over this on
> the weekend but I am not going to be able to rerun this test until
> some time in the week. You could enjoy the weekend and just look at
> this during the week.

Hi Konrad,
     We should have root-caused this bug, which is caused by 
for_each_pci_root_bus().
Current implementation doesn't support root bus deletion when walking 
PCI root
buses by for_each_pci_root_bus(). The reference counter 
(pci_bus->dev.knode_class.n_ref)
becomes zero after returning from pci_remove_root_bus(), so it triggers 
kref warnings
and double-free of klist_node object when we call 
pci_get_next_root_bus() to get the
next PCI root bus.
     So we will first revert to  list_for_each_entry_safe(bus, temp, 
&pci_root_buses, node)
and solve this issue in next version of for_each_pci_root_bus().

Regards!
Gerry

---
diff --git a/drivers/pci/xen-pcifront.c b/drivers/pci/xen-pcifront.c
index 6aa2c0f..6e577db 100644
--- a/drivers/pci/xen-pcifront.c
+++ b/drivers/pci/xen-pcifront.c
@@ -553,11 +553,11 @@ static int pcifront_rescan_root(struct 
pcifront_device *pdev,
 static void pcifront_free_roots(struct pcifront_device *pdev)
 {
        struct pcifront_sd *sd;
-       struct pci_bus *bus;
+       struct pci_bus *bus, *temp;

        dev_dbg(&pdev->xdev->dev, "cleaning up root buses\n");

-       for_each_pci_root_bus(bus) {
+       list_for_each_entry_safe(bus, temp, &pci_root_buses, node) {
                sd = bus->sysdata;
                if (sd->pdev == pdev) {
                        pci_stop_root_bus(bus);
---